Neural networks with Euclidean Symmetry for Learning from Physical Systems

Abstract:
To use machine learning to tackle challenges in the chemical and biological sciences, we need methods built to handle the “datatypes” of physical systems: geometry and the geometric tensors. These are traditionally challenging datatypes to use for machine learning because coordinates and coordinate systems are sensitive to the symmetries of 3D space: 3D
rotations, translations, and inversion.

In this talk, I present a method that I have been developing with my colleagues for the past five years, Euclidean neural networks. These networks preserve Euclidean symmetry by construction, making them incapable of unphysical bias due to a change of coordinates. They eliminate the need for data augmentation -- the 500-fold increase in brute-force training
necessary for a model to learn 3D patterns in arbitrary orientations. This makes them extremely data-efficient; they result in more accurate models and require less training data to do so, which is ideal for modeling from scientific data that is expensive, difficult to acquire, or highly-varied.

I describe how Euclidean neural networks work, demonstrate their effectiveness on a variety of real-world tasks, and introduce new capabilities my colleagues and I are developing with these methods. I also show how to efficiently and flexibly build equivariant models using our
open-source PyTorch package e3nn (https://e3nn.org).